Instance-Segmentation-with-Discriminative-Loss

This code is pytorch implementation of Semantic Instance Segmentation with a Discriminative Loss Function (https://arxiv.org/abs/1708.02551) with CVPPP-2017 Dataset (https://www.plant-phenotyping.org/CVPPP2017)

I slightly updated and revised the code of https://github.com/Wizaron/instance-segmentation-pytorch which does not match with python3. The updated content is as follows.

• compatibility with python3

• revised the dataset which does not use lmdb

• added other networks (UNet, UNet with CBAM, DeepLabV3)

  • unfortunately semantic & instance segmentation result of DeepLabV3 are very bad

• ReSeg with CoordConv(with r) is possible

My paper 'Leaf Instance Segmentation with Attention Based U-Net & Discriminative Loss' which utilizes this code was submitted for participation at 'Summer Annual conference of IEIE, 2022 (https://conf.theieie.org/2022s/)'

Notification

• You should execute python -m visdom.server before training

• train3.py -> pred_list2.py -> evaluation.ipynb

• I couldn't solve the error of training Stacked Recurrent Hourglass. The training does not proceed from 2 epochs due to errors in the back-propagation process.

• Updated(23/06/19)

  • If you use the latest version of pytorch (torch==2.1.1+cu121), please use train3_update.py & model_update.py instead of train3.py & model.py
    • .next() method was deprecated (line 257 of model.py)

UNet-CBAM

CBAM (Convolutional Block Attention Module)

Architecture

• Instance Counter : predicts normalized # of leaf instances
• Semantic Head : predicts semantic mask (f.g / b.g)
• Instance Head : predicts 32 dims embedding space which has to be clustered by K-Means

Results

CVPPP

Scores on validation subset (28 images)

Model	Loss	Mean SBD	Mean FG Dice	Dic
ReSeg	0.1425	0.8599	0.9669	0.7143
SegNet	0.1295	0.8645	0.9695	0.6071
UNet	0.1137	0.8656	0.9703	0.5713
UNet-CBAM	0.1031	0.8813	0.9708	0.7143

Sample Predictions (UNet-CBAM)

input / pred / GT

References

ReNet (used in ReSeg) : https://arxiv.org/abs/1505.00393

ReSeg : https://arxiv.org/abs/1511.07053

SegNet : https://arxiv.org/pdf/1511.00561.pdf

CBAM (Convolutional Block Attention Module) : https://arxiv.org/abs/1807.06521